This poster will describe the performance of models trained with Chemprop and in-house datasets and how they compare to commercial models.

We propose a method for determining the mechanism of action, also called target deconvolution, by cross-referencing drug perturbations with gene perturbations.

Contrasting using molecular similarities against ML in the form of hierarchical classifiers trained with positive, hard negatives and unlabeled data to predict enzyme commission numbers that can act on a given query molecule.

A framework for prediction probable impurity formations within an organic reaction.

No description.

Using graph embedding and neural networks to predict missing enzymatic links and characteristics of putative catalyzing enzymes.

We present a machine-intelligence approach to predict reaction products by integrating deep neural networks and probabilistic rule-based inference based on chemical knowledge, acheiving more than 90% accuracy on the USPTO dataset.

No description.

We present algorithms for automation friendly synthetic planning of linear natural products, optimization of  building block libraries and characterization of necessary cross couplings.

We propose a novel tree search algorithm for retrosynthetic planning which exploits neural modules. This enables the learning of generalizable structures from prior experiences to promote planning efficiency and effectiveness in new tasks.

No description. 

The ATOM Modeling PipeLine (AMPL) is an open source software library for building and sharing machine learning models that predict biological activities or physicochemical and pharmacokinetic properties.

We introduce conditional language models for protein sequences that directly condition on a graph specification of a 3D structure.

Can we integrate several AI models to provide a unified solution to accelerate the R&D process?

A method for the prediction of IR spectra from the molecular structure of novel molecules using a machine learning model.

A machine learning approach to rapidly optimize proteins with only small amounts of training data.

We propose to learn a molecular latent space organized by binding affinities, by adding binding affinities multitask prediction to a variational autoencoder. 

No description.

We present ChemBERT, a pre-trained language model based on BERT, which enables automated chemical reaction extraction. 

We over-express gene mutations in cultured cells and profile their phenotype using images to identify their impact in the development of cancer.

We use machine learning to predict the biochemical activity of a compound based on its chemical structure combined with gene expression and image-based profiles of cells exposed to it.

Discovering small molecules using mass spectral database search based on probabilistic modeling.

A machine learning toolbox for predicting ligand activity towards therapeutically important proteins. 

This work presents the application and assessment of deep learning software for the generation of de novo candidate antagonists of the NMDA receptor.

We use machine learning to predict the biochemical activity of a compound based on its chemical structure combined with gene expression and image-based profiles of cells exposed to it.

A deep generative model to enhance molecules with more desirable properties which utilizes a copy and refine strategy.

GNN model to predict major products for general regioselective reactions using quantum descriptors. 

We explore different techniques for property prediction of a molecular pair.

This poster presents an optimization strategy to plan syntheses for a molecular library.

We carefully benchmark property prediction models on public and proprietary datasets and introduce a new graph convolution model which outperforms previous baselines.

Work towards in-silico ligand design for Buchwald-Hartwig cross coupling reactions.

We identify several key challenges that need to be overcome for machine learning to be deployed in real-world drug discovery, propose a set of tools to address them and conclude with a case study highlighting the computational design of a potent compound inhibiting uptake of α-Synuclein by neuronal iPS cells with applications in Parkinson’s Disease.

Prediction of solubility using directed message passing neural networks.

We perform graph representation and generation by combining message passing neural networks and the optimal transport geometry.

In this work, we explore how to learn effective drug representation using BERT by leveraging the massive unlabeled molecular data and the bidirectional self-attention mechanisms for improved molecular property prediction.

The method to represent and generate 3D molecular structures using GAN will be presented.

We propose the Molecular Prediction Model Fine-Tuning (MolPMoFiT) approach, an effective transfer learning method that can be applied to any QSPR/QSAR problems.

We explore strategies for training text-based generative models of chemical structures in low-data situations, and apply these insights to train generative models of naturally occurring small molecules (metabolites) in several species.

A graph-to-graph translation model for optimizing molecular properties.

We will describe application and validation of Alchemite™, a novel deep learning method for data imputation, to drug discovery data covering two pharma projects and diverse endpoints.

This work takes a step towards an accurate characterization of DNN-based uncertainty for molecular property prediction, providing experimental insights based on both public and industry datasets.

To address the rapid emergence of antibiotic-resistant bacteria, we trained a deep neural network capable of predicting molecules with antibacterial activity.

Application of multi-armed bandit problems to ensemble FMO to efficiently propose promising idea compounds.

The overlap and performance of reaction datasets for the task of retrosynthetic planning are explored, alongside the potential to restrict their domain for specialized applications.

Transfer learning enables us an opportunity to explore the extrapolation area.

Standigm developed an AI de novo drug design workflow from target to lead compound which automates molecule generation, prioritization and optimization for further synthesis and testing.

We use episodic meta-learning to learn to extrapolate and increase applicability domain of QSAR models.

We propose a new hierarchical graph pooling layer that improves graph neural networks performance and interpretability on several molecular property prediction benchmarks.

We present a deep neural network (DNN) that not only outperforms DFT in predicting energies and electron densities of organic molecules, but also has a strong theoretical foundation.

New methodology for variational autoencoder.

This poster presents results of pre-training a reaction template relevance neural network on generalized template applicability.

Through retrospective analysis of existing reactivity data derived from high-throughput reaction screening, we demonstrate that active machine learning can be used to reduce the experimental burden associated with this type of screening.

A Weisfeiler-Lehman network based model was proposed to predict reaction conditions.

We developed an unsupervised learning approach to learn the dynamics of atoms or small molecules in materials from time-series molecular dynamics simulation data.

ATOM consortium developed a platform for generative molecular design an demonstrated its application in lead optimization by discovery of potent, selective aurora kinase B inhibitors with favorable candidate quality attributes.

Motivated by molecular optimization as a translation task, we develop Black Box Recursive Translation (BBRT), a drop-in framework where we explore and demonstrate the empirical benefits of recursive inference, vis-à-vis decoding strategies, as applied to unconstrained molecular optimization.

We present a method that improves the performance of small-molecule-ligand binding pose prediction by considering other ligands known to bind the target protein.

We use full, raw 3D electronic structure gas-phase quantum chemistry calculations as input to a convolutional neural network for the prediction of a variety of molecular and condensed phase material properties, including transfer learning to a small experimental dataset.

We present a method that improves the accuracy of small-molecule-ligand binding pose prediction by considering other ligands known to bind the target protein.

A message passing neural network predicts the activation energies of gas-phase chemical reactions.

An encoding of periodic trends is employed to develop a high-dimensional neural network potential which is constant in size with respect to the number of unique elements in the training data.

We present Olympus, a framework for the benchmarking of optimization algorithms on experimentally derived surfaces. Olympus includes a collection of datasets from physics, chemistry and materials science, and a suite of optimization algorithms, from random search to Bayesian and evolutionary, that can be easily accessed via a user-friendly python interface. In addition, Olympus allows to test custom algorithms and user-defined benchmark datasets. Here, we will present the user interface and a few examples of how Olympus can be used as a benchmark toolkit for optimization, as well as a simulation platform for hypotheses evaluation prior to experimental testing.

This poster describes a robust and automated approach for data analytics method selection and model construction that allows the user to focus on goals rather than methods.

Using a dataset generated from quantum chemistry, we predict 3D transition state structures from reactant and product geometries.

We show that Doc2Vec-generated embeddings for biomedical entities can be useful for nearest neighbor, clustering, and association learning tasks.

Iterative machine learning framework integrates multiple data sources to link small molecules to therapeutic targets.

We propose a sequential Monte Carlo algorithm to find diverse synthesis routes of target molecule within a Bayesian framework, and accelerated the sampling efficiency by a surrogate model.

We learn protein representations by integrating data from physical interaction and amino acid sequence.

A method for learning effective deep learning representations of promoter sequences with low resource data.

Uses of machine learning models and feature engineering will be presented for predicting Mab developability.

We quantified synthetic feasibility with a a data-driven computer-aided synthesis planning program and analyzed the approaches to bias generative models to synthesizable space.

Big data  genomics and machine learning can be used to understand T cell dysfunction in human tumors. We created an invitro system to improve our understanding of how T cells get exhausted in tumors and how we can use these data to identify new targets and biomarkers.

We present a machine learning methodology for lead optimization and scaffold hopping which combines synthesis prediction and uncertainty-calibrated bioactivity prediction.

Overview of the newly established Phase I CCI. 

Method to build the molecular structure one bond at a time given its chemical formula and observed spectrum.

The poster will introduce the SynSpace de novo design software with built-in synthetic feasibility and retrosynthesis capability that provides a user-friendly environment for drug hunters and a versatile design engine for automated platforms.

Enhancing process development via hybrid approach of using a prior knowledge and machine learning. 

Applying state-of-art deep learning models to design novel JAK inhibitors. 

Apply fragmentation and machine learning methods to rank the activities of Cathepsin S inhibitors which is challenging for conventional methods due to their large size, high flexibility and similar chemical component.

We present a deep neural network model for recognizing graph structures from chemical compound images.

Our drug discovery pipeline combines molecular docking and molecular dynamics simulations with powerful image processing and machine learning techniques.

Combining matched molecular pair analysis(MMPA) derived rules for ADMET issues with regression forest based potency scoring in an active learning mode can be used to accelerate lead optimization.

Iterative screening can lead to drastically improved efficiency in hit discovery compared to the traditional HTS paradigm.